feat: admit photos and image documents into the bound session

Telegram 图片消息此前一律回复「暂只支持文本消息」。本改动：

- photo 取最大尺寸（TG 压成 jpeg），document 仅接受 image/png|jpeg|webp|gif
- caption 作为伴随文本块进入同一条消息（不解析为命令）
- 相册（同一 media_group_id）去抖 1.2s 合并为一条多图消息，避免 N 连回合
- 派发走宿主 sessions.prompt RPC（Web 端同通道）：模型图像能力校验 +
  附件服务持久化 + mode queue 排队，与纯文本路径的 agent.followup 语义一致
- 错误映射：MODEL_DOES_NOT_SUPPORT_IMAGES → 提示换多模态模型；其余回退
  图片发送失败 + [code] 详情；下载失败即时提示
- 客户端新增 getFile/downloadFile（GET file 端点，复用代理感知 fetch）

tests +7（58/58）：photo+caption、png document、pdf 拒绝、相册合并、
模型不支持、下载失败、未绑定。无 emoji 文案。

lib/bridge.d.ts

@@ -53,6 +53,8 @@ export declare class TelegramBridge {
     private readonly compacting;
     /** sessionId → compaction abort controller (cancelled on bridge stop) */
     private readonly compactAborts;
+    /** Album accumulation: chatId:media_group_id → parts buffered into one prompt. */
+    private readonly mediaGroups;
     constructor(ctx: Context, options: TelegramBridgeOptions);
     /** Effective rendering mode for a chat: per-chat override, else the global mode. */
     private modeFor;
@@ -83,6 +85,28 @@ export declare class TelegramBridge {
     /** 等待压缩落定并汇报结果（不阻塞轮询循环）。 */
     private runCompaction;
     private followupBound;
+    /**
+     * Extract wire prompt parts from a media message.
+     * Returns undefined when the message carries no media at all; returns [] for
+     * an unsupported media type (the user was already notified); otherwise the
+     * caption text part (when present) followed by the image part.
+     */
+    private collectMediaParts;
+    /** Download one TG file and encode it as a canonical-base64 wire image part. */
+    private downloadImagePart;
+    /**
+     * Album photos arrive as separate updates sharing a media_group_id — debounce
+     * them into ONE prompt so an album does not spawn N sequential turns.
+     */
+    private accumulateMediaGroup;
+    /**
+     * Dispatch an image-bearing message to the bound session through the host
+     * prompt RPC — the exact channel the Web composer uses. The host checks
+     * whether the current model accepts image input, admits the bytes into the
+     * durable attachment store, then queues the message on the agent (mode
+     * 'queue' — same semantics as the text path's agent.followup).
+     */
+    private sendImageFollowup;
     private pollLoop;
     private interruptibleDelay;
     private interruptibleSleep;